A non-destructive testing method and system for underground void disease

Abstract

The application discloses a kind of nondestructive testing method and system for underground void disease, belong to the nondestructive testing technical field of road underground disease, method includes: through radar acquisition GPR original image, constructs GPR void image dataset;DIA-YOLO detection model is constructed;Through GPR void image dataset, DIA-YOLO detection model is trained, and void professional detection model is obtained;The GPR image to be detected is input into void professional detection model, and underground void disease detection result is obtained.The application realizes high-precision real-time detection of void disease by integrating deformable interactive attention module into YOLO11 backbone network and optimizing adaptive parameters, combined with hardware configuration, provides a more efficient solution for void investigation of road, tunnel and other projects, and solves the problems of poor detection accuracy, insufficient detection efficiency and radar image detection results lagging behind the acquisition process in the prior art.

Description

Technical Field

[0001] This invention relates to the field of non-destructive testing technology for underground road defects, specifically to a non-destructive testing method and system for underground void defects. Background Technology

[0002] Ground-penetrating radar (GPR) has become a core technology for non-destructive detection of road surface voids due to its significant advantages of non-destructive operation and high efficiency. Currently, mainstream GPR image detection algorithms include classic deep learning target detection algorithms such as YOLO11, YOLOv8, Faster R-CNN, and SSD. However, in scenarios with high real-time requirements, such as actual road inspection, current mainstream intelligent solutions have significant limitations. On the one hand, existing methods typically employ static feature extraction mechanisms with fixed receptive fields, making it difficult to adapt to the irregular shapes and weak signal characteristics of voids, resulting in high false negative and false positive rates. On the other hand, existing methods have a large number of model parameters, resulting in slow image detection speeds and drawbacks such as delayed radar image detection results and long operation cycles. In the entire process of data acquisition, detection, and localization in the field, they cannot quickly output the location, boundary, and confidence information of voids, making it difficult to provide accurate data support for the efficient maintenance and safety management of transportation infrastructure. Summary of the Invention

[0003] In view of the above-mentioned shortcomings in the prior art, the present invention provides a non-destructive testing method and system for underground void defects, which solves the problems of poor detection accuracy, insufficient detection efficiency, and radar image detection results lagging behind the acquisition process in the prior art.

[0004] To achieve the aforementioned objectives, the technical solution adopted by this invention is: a non-destructive testing method for underground void defects, comprising the following steps: S1. Acquire raw GPR images using radar, preprocess them, and construct a GPR de-gaussing image dataset; S2. Construct a DIA-YOLO detection model based on YOLO11. The specific process is as follows: replace the C2PSA module after the SPPF module in the backbone network with a deformable interactive attention module. S3. Train the DIA-YOLO detection model using the GPR declassified image dataset to obtain a professional declassification detection model; S4. Input the GPR image to be detected into the underground void detection model to obtain the detection results of underground void disease.

[0005] Furthermore: In S1, the GPR delabeled image dataset includes an initial labeled dataset and an unlabeled dataset. The method for constructing the GPR delabeled image dataset includes the following steps: S11. Non-destructive scanning of underground structures is performed using ground-penetrating radar to collect raw GPR images containing underground voids and construct an initial unlabeled dataset. S12. Preprocess the initial unlabeled dataset, manually label the preprocessed GPR vacuolated images, and divide the labeling categories into vacuolated diseases and underground structures to construct the initial labeled dataset. S13. Enhance the preprocessed GPR de-voided images to target the underground de-voided disease characteristics, and obtain an unlabeled dataset.

[0006] Furthermore, in S12, the preprocessing methods include: S121. Remove DC; after removing DC, the GPR image shows depth... Horizontal pixel value at The specific expression is: In the formula, For the original GPR image in depth Horizontal Pixel value at that location, This represents the total number of horizontal trajectories. S122. Time zero-point correction: After time zero-point correction, the GPR image shows depth... Horizontal pixel value at The specific expression is: In the formula, The offset for zero-point time correction. Total pixels in the depth direction; For modulo operation, To remove DC from the GPR image in depth Horizontal Pixel value at; S123. Automatic gain control; after automatic gain control, the GPR image shows depth... Horizontal pixel value at The specific expression is: In the formula, For the first The local root mean square of a depth window. The target is the root mean square. In the formula, It is the minimum value. Local energy; In the formula, N To control the depth direction of the sliding window using automatic gain control, For the GPR image after time zero-point correction in depth Horizontal pixel value at The filled pixel values; S124. Background removal; after background removal, the GPR image shows depth... Horizontal pixel value at The specific expression is: In the formula, Total pixels in the depth direction; S125, Lateral averaging, after lateral averaging, the GPR image shows depth... Horizontal pixel value at The specific expression is: In the formula, After background removal, the GPR image has depth Horizontal Pixel value at; S126, Minimum-Max Normalization, after minimum-maximum normalization, the GPR image in depth Horizontal pixel value at The specific expression is: In the formula, for The minimum pixel value, for The maximum pixel value.

[0007] Furthermore, in S13, methods to address the increased severity of underground voiding include: S131. Gaussian noise injection. The specific expression for Gaussian noise injection is as follows: In the formula, The maximum pixel value of the normalized image. For signal-to-noise ratio, The standard deviation of noise; S132, Directional Brightness / Contrast Adjustment, is used to increase the grayscale difference between the weak void signal and the background, making it easier for the model to identify the blurred void edges; Brightness adjustment, the specific expression for brightness adjustment is: In the formula, For normalized images, The image after brightness adjustment; Contrast adjustment, the specific expression for contrast adjustment is: In the formula, This is the image after contrast adjustment. For normalized images Transformation, For coefficients, Randomly select 0.9 or 1.1; S133, Affine transformations specific to the empty region, which include scaling, rotation and translation operations; S134, Constrained Mosaic Enhancement: Randomly select 4 images, of which 2 with gaps are stitched together to form a 640×640 pixel image.

[0008] Furthermore: In S2, the DIA-YOLO detection model includes a backbone network, a neck network, and a head network connected in sequence; The backbone network comprises a first convolutional module, a second convolutional module, a first C3K2 module, a third convolutional module, a second C3K2 module, a fourth convolutional module, a third C3K2 module, a fifth convolutional module, a fourth C3K2 module, an SPFF module, and a deformable interactive attention module, all connected in sequence. The neck network comprises a first Upsample module, a first Concat module, a fifth C3K2 module, a second Upsample module, a second Concat module, a sixth C3K2 module, a sixth convolutional module, a third Concat module, a seventh C3K2 module, a seventh convolutional module, a fourth Concat module, and an eighth C3K2 module, connected sequentially. The input of the first Upsample module is connected to the output of the deformable interactive attention module; the input of the first Concat module is also connected to the output of the third C3K2 module; the output of the fifth C3K2 module is also connected to the input of the third Concat module; the input of the second Concat module is also connected to the output of the second C3K2 module; and the input of the fourth Concat module is also connected to the output of the deformable interactive attention module. The head network includes a first detection head, a second detection head, and a third detection head. The input of the first detection head is connected to the output of the sixth C3K2 module, the input of the second detection head is connected to the output of the seventh C3K2 module, and the input of the third detection head is connected to the output of the eighth C3K2 module.

[0009] Furthermore: In S2, the workflow of the deformable interactive attention module is as follows: A1. De-emptying features for input deformable interactive attention modules Through global average pooling The first feature is obtained by capturing the global contour of the void from the void-free features. Through global max pooling Extract the weak edge features of the void-free feature to obtain the second feature. ; A2. Regarding the first feature The third feature is obtained by sequentially performing downsampling, contour modulation, gradient scaling, and upsampling. The contour feature weights of contour modulation are: The gradient scaling factor is ; A3. Regarding the second feature The fourth feature is obtained by sequentially performing downsampling, edge modulation, gradient scaling, and upsampling. The edge feature weights of edge modulation are: The gradient scaling factor is ; A3. The third feature and the fourth feature The layers are concatenated, and a spatial mask is generated using a 1×1 convolution and the Sigmoid function. ; A4. Through spatial masking With emptying characteristics Perform element-wise feature multiplication and output the enhanced empty features.

[0010] Furthermore, in S2, the head network of the DIA-YOLO detection model assigns a dedicated category channel to the void disease, adopts an adaptive boundary focus loss function, sets the void target loss weight to 0.8 and the background loss weight to 0.2, and strengthens the model's learning priority for weak void signals.

[0011] Furthermore: In S3, the specific method for training the DIA-YOLO detection model is as follows: S31. Input the initial labeled dataset into the YOLO11X detection model and start the teacher model training: use the Adam optimizer, with a batch size of 16 and 100 iterations. Combine the adaptive boundary focus loss function to optimize the model parameters. At the same time, set gradient clipping to suppress gradient explosion and set the EMA decay coefficient to 0.9999 until the model's mAP@0.5 tends to converge, and the teacher model is obtained. S32. Augmenting the unlabeled dataset using a teacher model: The teacher model performs inference on the unlabeled dataset. For each augmented GPR de-defined image, the teacher model outputs the predicted bounding box, category label, and confidence score of the de-defined region. High-confidence prediction results with a confidence score greater than or equal to 0.75 are selected as initial pseudo-labels. For the initial pseudo-labels, 25% of the pseudo-label samples are manually sampled to correct recognition errors, border offset annotations, and missing annotations. At the same time, unstable annotations with large confidence fluctuations are automatically filtered out to form a high-quality augmented labeled dataset. S33. Integrate and divide the dataset: merge the initial labeled dataset and the enhanced labeled dataset. After merging, adopt a hierarchical random partitioning strategy to split the dataset into training set and test set in a 7:3 ratio, so that the data in the enhanced labeled dataset and the data before enhancement are assigned to the same set, and control the proportion of vacant disease in the training set and test set to be consistent. S34. Using the DIA-YOLO detection model as the student model, input the training set and test set into the student model to start the training of the blank detection model: use the core training parameters of the teacher model, and introduce a learning rate decay strategy. During the training process, monitor the precision, recall and mAP@0.5 of the test set in real time; when the test set metrics do not show significant improvement for 50 consecutive iterations, stop training, save the current optimal model weights, and obtain the blank detection professional model.

[0012] A non-destructive testing system for underground void defects, the system comprising: The radar data acquisition equipment includes an antenna array, an encoder, and a radar host, with both the antenna array and the encoder connected to the radar host. The computing unit is connected to the radar host and deploys the DIA-YOLO detection model; Data storage and display equipment, including a display screen and a mobile storage device, both of which are connected to a computing unit.

[0013] The beneficial effects of this invention are as follows: (1) This invention provides a non-destructive testing method for underground void defects. By integrating the deformable interactive attention module into the YOLO11 backbone network and optimizing the adaptation parameters, combined with hardware configuration, high-precision real-time detection of void defects is achieved, providing a more efficient solution for void investigation in road, tunnel and other engineering projects.

[0014] (2) This invention constructs a non-destructive testing system for underground void defects. Compared with other intelligent detection systems for void defects, this invention can achieve real-time detection of void defects. The detection results are output synchronously with the data acquisition process. Operators can stop their vehicles immediately upon discovering void defects and conduct core sampling verification on-site based on precise positioning information, avoiding problems such as vehicles leaving or positioning deviations due to detection delays. This significantly improves operational efficiency and meets the practical engineering requirements of simultaneous data acquisition, detection, and verification.

[0015] In summary, the real-time non-destructive testing method for underground voiding defects based on the DIA-YOLO detection model proposed in this invention specifically addresses the problems of poor detection accuracy, insufficient detection efficiency, and the lag of radar image detection results behind the acquisition process in existing technologies. Attached Figure Description

[0016] Figure 1 This is a flowchart of a non-destructive testing method for underground void defects according to the present invention.

[0017] Figure 2 This is a structural diagram of the DIA-YOLO detection model.

[0018] Figure 3 This is a schematic diagram of a deformable interactive attention module.

[0019] Figure 4 This is a diagram showing the results of the delamination test.

[0020] Figure 5 This is a schematic diagram of a non-destructive testing system for underground voiding defects according to the present invention.

[0021] Figure 6 This is a schematic diagram of a vehicle-mounted deployment device.

[0022] Figure 7 This is a schematic diagram for detecting road surface defects. Detailed Implementation

[0023] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0024] like Figure 1 As shown, in one embodiment of the present invention, a non-destructive testing method for underground void defects includes the following steps: S1. Acquire raw GPR images using radar, preprocess them, and construct a GPR de-gaussing image dataset; S2. Construct a DIA-YOLO detection model based on YOLO11. The specific process is as follows: replace the C2PSA module after the SPPF module in the backbone network with a deformable interactive attention module (DIA-Module). S3. Train the DIA-YOLO detection model using the GPR declassified image dataset to obtain a professional declassification detection model; S4. Input the GPR image to be detected into the underground void detection model to obtain the detection results of underground void disease.

[0025] In S1, the GPR delabeled image dataset includes an initial labeled dataset and an unlabeled dataset. The method for constructing the GPR delabeled image dataset includes the following steps: S11. Non-destructive scanning of underground structures is performed using ground-penetrating radar to collect raw GPR images containing underground voids and construct an initial unlabeled dataset. S12. Preprocess the initial unlabeled dataset, manually label the preprocessed GPR vacuolated images, and divide the labeling categories into vacuolated diseases and underground structures to construct the initial labeled dataset. S13. Enhance the preprocessed GPR de-voided images to target the underground de-voided disease characteristics, and obtain an unlabeled dataset.

[0026] In S12, the preprocessing methods include: S121. Remove DC to eliminate DC offset caused by hardware system or antenna coupling or ground reflection, making the signal centered at zero to highlight weak signal differences. After removing DC, the GPR image shows depth... Horizontal pixel value at The specific expression is: In the formula, For the original GPR image in depth Horizontal Pixel value at that location, This represents the total number of horizontal trajectories. S122. Time zero-point correction corrects zero-point misalignment in GPR images caused by hardware delays, aligning the gap signals at different lateral positions. After time zero-point correction, the GPR image shows improved depth accuracy. Horizontal pixel value at The specific expression is: In the formula, The offset for zero-point time correction. Total pixels in the depth direction; For modulo operation, To remove DC from the GPR image in depth Horizontal Pixel value at; S123. Automatic gain control compensates for energy attenuation caused by GPR image propagation in the medium, amplifying weak signals from deep voids. Local energy is calculated using a sliding window and normalized to the target root mean square (RMS) to ensure signal energy consistency across different depths, thus amplifying the weak signals from deep voids. After automatic gain control, the GPR image at different depths... Horizontal pixel value at The specific expression is: In the formula, For the first The local root mean square of a depth window. The target is the root mean square. In the formula, To find the minimum value, to prevent the denominator from being 0. Local energy; In the formula, N To control the depth direction of the sliding window using automatic gain control, For the GPR image after time zero-point correction in depth Horizontal pixel value at The filled pixel values; S124. Background Removal: This removes background noise amplification caused by gain, further highlighting the detached target echo. After background removal, the GPR image shows improved depth. Horizontal pixel value at The specific expression is: In the formula, Total pixels in the depth direction; S125. Lateral averaging: Averaging two adjacent lateral trajectories suppresses lateral random noise while preserving lateral spatial resolution. After lateral averaging, the GPR image has improved depth... Horizontal pixel value at The specific expression is: In the formula, After background removal, the GPR image has depth Horizontal Pixel value at; S126. Minimum-maximum normalization linearly maps the data to the [0,1] interval, facilitating subsequent real-time visualization and model processing. After minimum-maximum normalization, the GPR image shows depth... Horizontal pixel value at The specific expression is: In the formula, for The minimum pixel value, for The maximum pixel value.

[0027] In S13, the methods to address the increased severity of underground voiding include: S131. Gaussian noise injection: Gaussian noise with a mean of 0 and a signal-to-noise ratio of 40±10dB is added to simulate electromagnetic clutter in the actual engineering environment, enhancing the model's anti-interference capability. The specific expression for Gaussian noise injection is as follows: In the formula, The maximum pixel value of the normalized image. For signal-to-noise ratio, The standard deviation of noise; S132, Directional Brightness / Contrast Adjustment, is used to increase the grayscale difference between the weak void signal and the background, making it easier for the model to identify the blurred void edges; Brightness adjustment: With a fixed linear transformation coefficient of ±10%, the balance between weak signal amplification and noise suppression is achieved. The specific expression for brightness adjustment is as follows: In the formula, For normalized images, The image after brightness adjustment; Contrast adjustment is used to amplify the grayscale difference between weak signals and the background. The specific expression for contrast adjustment is: In the formula, This is the image after contrast adjustment. For normalized images Transformation, For coefficients, Randomly select 0.9 or 1.1; S133. Affine transformations specific to voided regions: Addressing the variable dimensions of voids, fixed-range geometric transformations are performed on the voided annotation regions to increase the diversity of void shapes without destroying their morphological characteristics. Methods for affine transformations specific to voided regions include: Scaling by 0.9 to 1.1 times, balancing the integrity and diversity of the hollowed-out form; Rotate by -5° to 5° to match the linear scanning characteristics of GPR and avoid excessive rotation that could damage the gap-free features; Shift by ±3 pixels to meet the precision requirements of 640×640 pixel input; S134. Constrained Mosaic Enhancement: Randomly select 4 images, of which 2 contain voids and stitch them together to form a 640×640 pixel image, ensuring that the void annotation boxes do not overlap and that the void feature is intact.

[0028] After applying the above four data augmentation methods, we obtain the data-augmented unlabeled dataset.

[0029] like Figure 2 As shown in Figure S2, the DIA-YOLO detection model includes a backbone network, a neck network, and a head network connected in sequence. The backbone network comprises a first convolutional module, a second convolutional module, a first C3K2 module, a third convolutional module, a second C3K2 module, a fourth convolutional module, a third C3K2 module, a fifth convolutional module, a fourth C3K2 module, an SPFF module, and a deformable interactive attention module, all connected in sequence. The neck network comprises a first Upsample module, a first Concat module, a fifth C3K2 module, a second Upsample module, a second Concat module, a sixth C3K2 module, a sixth convolutional module, a third Concat module, a seventh C3K2 module, a seventh convolutional module, a fourth Concat module, and an eighth C3K2 module, connected sequentially. The input of the first Upsample module is connected to the output of the deformable interactive attention module; the input of the first Concat module is also connected to the output of the third C3K2 module; the output of the fifth C3K2 module is also connected to the input of the third Concat module; the input of the second Concat module is also connected to the output of the second C3K2 module; and the input of the fourth Concat module is also connected to the output of the deformable interactive attention module. The head network includes a first detection head, a second detection head, and a third detection head. The input of the first detection head is connected to the output of the sixth C3K2 module, the input of the second detection head is connected to the output of the seventh C3K2 module, and the input of the third detection head is connected to the output of the eighth C3K2 module.

[0030] In this embodiment, the workflow of the DIA-YOLO detection model is as follows: S21. Input the 640×640×3 GPR de-voiced image into the neck network. It is downsampled to 320×320× the corresponding number of channels by a convolutional module with a 3×3 kernel and a stride of 2. Basic gray-level edge features are extracted and high-frequency clutter is filtered out. The convolutional features are fed into the C3K2 module (residual branch + group convolution) to discover the weak gray-level patterns of de-voiced features and suppress background clutter. Through the cyclical cooperation between the convolutional module and the C3K2 module, the convolutional module halves the feature map size (e.g., 640→320→160) and increases the number of channels. The C3K2 module refines de-voiced features of different sizes at the new scale. After multiple rounds of processing, the feature map is input into the SPPF module, where it is pooled in parallel using 3×3 / 5×5 / 7×7 pooling kernels to stitch features together to preserve small void details, cover large void contours, and adapt to the irregularity of void shapes. The void-clearing features output by the SPPF module are directly fed into the deformable interactive attention module. The workflow of the deformable interactive attention module is as follows: First, the void-clearing contour features are extracted through global average pooling, and the void-clearing edge features are captured through global max pooling. Then, a 3×3 convolutional layer with a stride of 2 is used to downsample these two types of features, and then the void-clearing signal is amplified by weighting the contour features with a weight of 0.3 and the edge features with a weight of 0.5.

[0031] Bilinear interpolation upsamples the data to the output size of the SPPF module and stitches them together. Then, it is compressed to a single channel through 1×1 convolution and a spatial mask in the range of 0 to 1 is generated by the Sigmoid function. Regions with a mask value of 1 correspond to the void removal feature, and regions with a mask value of 0 correspond to background clutter. The spatial mask is multiplied element-wise with the void removal feature to enhance the void removal feature and shield the clutter, outputting the enhanced void removal feature. S22. The enhanced void-free features are input into the neck network. First, an upsampling operation is used to enlarge the small-sized feature maps in high-level layers to the size of mid-level feature maps. Then, a concat operation is used to concatenate them with mid-level feature maps from a certain layer in the backbone network, fusing the void-free category attributes of high-level features with the void-free boundary positions of mid-level features. The concatenated features are then further refined in the C3K2 module. After the above iterations, a Conv downsampling operation is used to fuse shallow features, ultimately constructing a large / medium / small-scale feature pyramid to cover voids of different sizes.

[0032] S23. Three detection heads are respectively connected to feature maps of different scales. The first detection head identifies large gaps, the second detection head identifies regular gaps, and the third detection head identifies small gaps. Each detection head outputs three types of information: classification, location parameters, and confidence level.

[0033] Finally, the raw results output by the detection head undergo a post-processing process. First, the bounding box coordinates are decoded, then low-confidence predicted boxes with a confidence level below 0.25 are filtered out. Next, duplicate predicted boxes for the same empty region are removed by non-maximum suppression IoU with a threshold of 0.7. Finally, the center position, width and height dimensions, and confidence level of each empty region are output to complete the entire GPR empty image detection process.

[0034] like Figure 3 As shown, the workflow of the deformable interactive attention module in S2 is as follows: A1. De-emptying features for input deformable interactive attention modules Through global average pooling The first feature is obtained by capturing the global contour of the void from the void-free features. Through global max pooling Extract the weak edge features of the void-free feature to obtain the second feature. ; A2. Regarding the first feature The third feature is obtained by sequentially performing downsampling, contour modulation, gradient scaling, and upsampling. The contour feature weights of contour modulation are: The gradient scaling factor is ; A3. Regarding the second feature The fourth feature is obtained by sequentially performing downsampling, edge modulation, gradient scaling, and upsampling. The edge feature weights of edge modulation are: The gradient scaling factor is ; A3. The third feature and the fourth feature The layers are concatenated, and a spatial mask is generated using a 1×1 convolution and the Sigmoid function. ; A4. Through spatial masking With emptying characteristics Perform element-wise feature multiplication and output the enhanced empty features.

[0035] The head network of the DIA-YOLO detection model assigns a dedicated category channel to void defects, adopts an adaptive boundary focus loss function, sets the void target loss weight to 0.8 and the background loss weight to 0.2, and strengthens the model's learning priority for weak void signals.

[0036] In this embodiment, the adaptive boundary focus loss function The specific expression is: In the formula, For classifying losses, The bounding box regression loss is (CIOU loss). For classification loss weights, =0.8, Weights for the bounding box loss. =0.2, In S3, the specific method for training the DIA-YOLO detection model is as follows: S31. Input the initial labeled dataset into the YOLO11X detection model and start the teacher model training: use the Adam optimizer, with a batch size of 16 and 100 iterations. Combine the adaptive boundary focus loss function to optimize the model parameters. At the same time, set gradient clipping to suppress gradient explosion and set the EMA decay coefficient to 0.9999 to improve model stability until the model's mAP@0.5 tends to converge, and the teacher model is obtained. S32. Augmenting the unlabeled dataset using a teacher model: The teacher model performs inference on the unlabeled dataset. For each augmented GPR de-defined image, the teacher model outputs the predicted bounding box, category label, and confidence score of the de-defined region. High-confidence prediction results with a confidence score greater than or equal to 0.75 are selected as initial pseudo-labels. For the initial pseudo-labels, 25% of the pseudo-label samples are manually sampled to correct recognition errors, border offset annotations, and missing annotations. At the same time, unstable annotations with large confidence fluctuations are automatically filtered out to form a high-quality augmented labeled dataset. S33. Dataset integration and partitioning: The initial labeled dataset and the augmented labeled dataset are merged. After merging, a hierarchical random partitioning strategy is adopted to split the dataset into training and test sets in a 7:3 ratio. This ensures that the data in the augmented labeled dataset is placed in the same set as the data before augmentation, and that the proportion of vacancy diseases in the training and test sets is consistent. This improves the generalization ability of the model while preventing the test results from being artificially inflated. S34. Using the DIA-YOLO detection model as the student model, input the training set and test set into the student model to start the training of the blank detection model: use the core training parameters of the teacher model, and introduce a learning rate decay strategy. During the training process, monitor the precision, recall and mAP@0.5 of the test set in real time; when the test set metrics do not show significant improvement for 50 consecutive iterations, stop training, save the current optimal model weights, and obtain the blank detection professional model.

[0037] After completing the above training steps, this invention constructs a professional void detection model, ensuring that the void detection accuracy meets engineering requirements, such as... Figure 4 As shown, Figure 4In the table, Viod 0.8 indicates a detection accuracy of 0.8 for missing data. Compared to YOLO11x, which is used as the teacher model, it demonstrates a significant lightweight advantage, as shown in Table 1. This provides reliable support for engineering practices involving simultaneous data acquisition, detection, and on-site verification.

[0038] Table 1 Model Comparison Table like Figure 5 As shown, a non-destructive testing system for underground void defects includes: The radar data acquisition equipment includes an antenna array (DXG1820), an encoder, and a radar host. Both the antenna array and the encoder are connected to the radar host. The computing unit is connected to the radar host and deploys the DIA-YOLO detection model; Data storage and display equipment, including a display screen and a mobile storage device, both of which are connected to a computing unit.

[0039] like Figure 6 As shown, in this embodiment, a non-destructive testing system for underground void defects is deployed in a vehicle-mounted deployment device. The vehicle-mounted deployment device includes a testing vehicle, a trailer-mounted tow frame, and an acrylic plate. The trailer-mounted tow frame stably connects the testing vehicle and the radar probe, ensuring that the two operate synchronously and avoiding errors caused by differences in speed and path. The acrylic plate is used to isolate the radar probe from the road surface, preventing damage to the radar probe and reducing clutter.

[0040] The antenna array employs a towable design, secured to the rear of the detection vehicle via a trailer-mounted tow rack. Operating at frequencies from 100MHz to 3000MHz, it emits electromagnetic waves and receives echoes reflected from underground structures. The encoder connects to the radar unit, acquiring real-time vehicle speed data to ensure synchronization between the GPR image and the vehicle's trajectory. The radar unit, located inside the detection vehicle, connects to the encoder via a custom interface, receiving data and generating GPR images. A high-speed communication link is established by the recomputation unit.

[0041] The process of detecting road void defects using the system of this invention is as follows: Figure 7 As shown, after the test vehicle is started, it drags the radar traction bracket at a constant speed. The radar emits electromagnetic waves to the ground through the signal transmitter. The reflected signals from the ground enter the receiver and are transmitted to the radar host via the encoder. The radar host and the acquisition computer are connected via an RJ45 network port, and the images are finally displayed on the computer as GPR images. The acquired GPR images are then transmitted to a dedicated void detection system to detect voids in real time, outputting the void location border and void confidence level. For areas detected as voids, on-site core sampling is immediately performed for verification. The system's workflow is as follows: Step 1: Before operation, fix the ground-penetrating radar to the rear of the detection vehicle using a tow bracket. After installing an acrylic plate under the probe, adjust the probe height to make it close to the road surface. Connect the radar host to the encoder on top of the radar through a custom interface, and then establish a communication line between the radar host and the computing unit using an RJ45 network port. Step Two: The operator drives the testing vehicle to the starting point of the road section to be inspected, demarcates the temporary work area, and places warning cones along the boundary. Based on the structure of the road section to be inspected and the required detection depth, parameters such as scanning interval and electromagnetic wave repetition frequency are set in the radar host. Data inference parameters are pre-set in the computing unit to ensure the algorithm runs smoothly for the road section to be inspected. The operator sends test commands through the computing unit and simultaneously connects the GPS positioning module and data acquisition system to check for any delays.

[0042] Step 3: After starting all equipment, the driver drives the test vehicle at a constant speed on the test section. The radar emits electromagnetic waves and receives reflected echoes from the ground. Data is synchronously transmitted to the radar host via the encoder, and the radar host transmits the data to the computing unit in the form of GPR images. The computing unit performs a customized preprocessing process on the original GPR images. First, DC offset is eliminated by DC removal, then depth alignment is achieved through time zero-point correction, then weak signals are amplified by AGC gain compensation, then background removal and two-stage average noise suppression are performed, and finally, the data is mapped to the [0,1] interval by max-min normalization to obtain a clear image of the void features after preprocessing. Step 4: Deploy the detachment detection model to the computing unit and establish a real-time data transmission connection with the ground-penetrating radar. Detected GPR images are input into the detachment detection model in real time. The computing unit quickly performs 640×640 scaling and 0~1 range normalization, performing real-time detection and accurately outputting the location bounding box, confidence score, and corresponding GPS coordinates of the detached target. The display screen shows the detection results in real time, marking the location and boundary of the detached road segment on the map. This achieves real-time acquisition, detection, and location without waiting for the entire road segment to be detected.

[0043] Step 5: Once the system detects and locates the voided target in real time, the operator can immediately park the testing vehicle according to the on-site work plan. Based on the precise GPS coordinates and voided boundary information provided by the system, the sampling location can be quickly locked. On-site core drilling and sampling verification can be carried out to directly verify the actual situation of the voided defect. There is no need to blindly cover the testing area, realizing a highly efficient operation mode of verifying voids as soon as they are discovered, greatly improving the targeting and efficiency of verification.

[0044] Step Six: After the inspection of the road section to be inspected is completed, the inspection vehicle drives to a safe area to carry out the final work. For missed or false detection cases that occur during real-time detection, these are treated as unlabeled datasets. Then, through data augmentation, pseudo-labels are generated using a teacher model, and the labels are manually corrected to expand the dataset. The unlabeled dedicated detection model is then retrained to improve its generalization ability.

[0045] In the description of this invention, the above are merely preferred embodiments and are not intended to limit the scope of protection of this invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A non-destructive testing method for underground void defects, characterized in that, Includes the following steps: S1. Acquire raw GPR images using radar, preprocess them, and construct a GPR de-gaussing image dataset; S2. Construct a DIA-YOLO detection model based on YOLO11. The specific process is as follows: replace the C2PSA module after the SPPF module in the backbone network with a deformable interactive attention module. S3. Train the DIA-YOLO detection model using the GPR declassified image dataset to obtain a professional declassification detection model; S4. Input the GPR image to be detected into the underground void detection model to obtain the detection results of underground void disease; In S2, the DIA-YOLO detection model includes a backbone network, a neck network, and a head network connected in sequence. The backbone network comprises a first convolutional module, a second convolutional module, a first C3K2 module, a third convolutional module, a second C3K2 module, a fourth convolutional module, a third C3K2 module, a fifth convolutional module, a fourth C3K2 module, an SPFF module, and a deformable interactive attention module, all connected in sequence. The neck network comprises a first Upsample module, a first Concat module, a fifth C3K2 module, a second Upsample module, a second Concat module, a sixth C3K2 module, a sixth convolutional module, a third Concat module, a seventh C3K2 module, a seventh convolutional module, a fourth Concat module, and an eighth C3K2 module, connected sequentially. The input of the first Upsample module is connected to the output of the deformable interactive attention module; the input of the first Concat module is also connected to the output of the third C3K2 module; the output of the fifth C3K2 module is also connected to the input of the third Concat module; the input of the second Concat module is also connected to the output of the second C3K2 module; and the input of the fourth Concat module is also connected to the output of the deformable interactive attention module. The head network includes a first detection head, a second detection head, and a third detection head. The input of the first detection head is connected to the output of the sixth C3K2 module, the input of the second detection head is connected to the output of the seventh C3K2 module, and the input of the third detection head is connected to the output of the eighth C3K2 module. In S2, the workflow of the deformable interactive attention module is as follows: A1. De-emptying features for input deformable interactive attention modules Through global average pooling The first feature is obtained by capturing the global contour of the void from the void-free features. Through global max pooling Extract the weak edge features of the void-free feature to obtain the second feature. ; A2. Regarding the first feature The third feature is obtained by sequentially performing downsampling, contour modulation, gradient scaling, and upsampling. The contour feature weights of contour modulation are: The gradient scaling factor is ; A3. Regarding the second feature The fourth feature is obtained by sequentially performing downsampling, edge modulation, gradient scaling, and upsampling. The edge feature weights of edge modulation are: The gradient scaling factor is ; A3. The third feature and the fourth feature The layers are concatenated, and a spatial mask is generated using a 1×1 convolution and the Sigmoid function. ; A4. Through spatial masking With emptying characteristics Perform element-wise feature multiplication and output the enhanced empty features.

2. The non-destructive testing method for underground void defects according to claim 1, characterized in that, In S1, the GPR delabeled image dataset includes an initial labeled dataset and an unlabeled dataset. The method for constructing the GPR delabeled image dataset includes the following steps: S11. Non-destructive scanning of underground structures is performed using ground-penetrating radar to collect raw GPR images containing underground voids and construct an initial unlabeled dataset. S12. Preprocess the initial unlabeled dataset, manually label the preprocessed GPR vacuolated images, and divide the labeling categories into vacuolated diseases and underground structures to construct the initial labeled dataset. S13. Enhance the preprocessed GPR de-voided images to target the underground de-voided disease characteristics, and obtain an unlabeled dataset.

3. The non-destructive testing method for underground void defects according to claim 2, characterized in that, In S12, the preprocessing methods include: S121. Remove DC; after removing DC, the GPR image shows depth... Horizontal pixel value at The specific expression is: In the formula, For the original GPR image in depth Horizontal Pixel value at that location, This represents the total number of horizontal trajectories. S122. Time zero-point correction: After time zero-point correction, the GPR image shows depth... Horizontal pixel value at The specific expression is: In the formula, The offset for zero-point time correction. The total number of pixels in the depth direction; For modulo operation, To remove DC from the GPR image in depth Horizontal Pixel value at; S123. Automatic gain control; after automatic gain control, the GPR image shows depth... Horizontal pixel value at The specific expression is: In the formula, For the first The local root mean square of a depth window. The target is the root mean square. In the formula, It is the minimum value. Local energy; In the formula, N To control the depth direction of the sliding window using automatic gain control, For the GPR image after time zero-point correction in depth Horizontal pixel value at The filled pixel values; S124. Background removal; after background removal, the GPR image shows depth... Horizontal pixel value at The specific expression is: In the formula, The total number of pixels in the depth direction; S125, Lateral averaging, after lateral averaging, the GPR image shows depth... Horizontal pixel value at The specific expression is: In the formula, After background removal, the GPR image has depth Horizontal Pixel value at; S126, Minimum-Max Normalization, after minimum-maximum normalization, the GPR image in depth Horizontal pixel value at The specific expression is: In the formula, for The minimum pixel value, for The maximum pixel value.

4. The non-destructive testing method for underground void defects according to claim 2, characterized in that, In S13, the methods to address the increased severity of underground voiding include: S131. Gaussian noise injection. The specific expression for Gaussian noise injection is as follows: In the formula, The maximum pixel value of the normalized image. For signal-to-noise ratio, The standard deviation of noise; S132, Directional Brightness / Contrast Adjustment, is used to increase the grayscale difference between the weak void signal and the background, making it easier for the model to identify the blurred void edges; Brightness adjustment, the specific expression for brightness adjustment is: In the formula, For normalized images, The image after brightness adjustment; Contrast adjustment, the specific expression for contrast adjustment is: In the formula, This is the image after contrast adjustment. For normalized images Transformation, For coefficients, Randomly select 0.9 or 1.1; S133, Affine transformations specific to the empty region, which include scaling, rotation and translation operations; S134, Constrained Mosaic Enhancement: Randomly select 4 images, of which 2 with gaps are stitched together to form a 640×640 pixel image.

5. The non-destructive testing method for underground void defects according to claim 2, characterized in that, In S2, the head network of the DIA-YOLO detection model assigns a dedicated category channel to the void disease, adopts an adaptive boundary focus loss function, sets the void target loss weight to 0.8 and the background loss weight to 0.2, and strengthens the model's learning priority for weak void signals.

6. The non-destructive testing method for underground void defects according to claim 2, characterized in that, In S3, the specific method for training the DIA-YOLO detection model is as follows: S31. Input the initial labeled dataset into the YOLO11X detection model and start the teacher model training: use the Adam optimizer, with a batch size of 16 and 100 iterations. Combine the adaptive boundary focus loss function to optimize the model parameters. At the same time, set gradient clipping to suppress gradient explosion and set the EMA decay coefficient to 0.9999 until the model's mAP@0.5 tends to converge, and the teacher model is obtained. S32. Augmenting the unlabeled dataset using a teacher model: The teacher model performs inference on the unlabeled dataset. For each augmented GPR de-defined image, the teacher model outputs the predicted bounding box, category label, and confidence score of the de-defined region. High-confidence prediction results with a confidence score greater than or equal to 0.75 are selected as initial pseudo-labels. For the initial pseudo-labels, 25% of the pseudo-label samples are manually sampled to correct recognition errors, border offset annotations, and missing annotations. At the same time, unstable annotations with large confidence fluctuations are automatically filtered out to form a high-quality augmented labeled dataset. S33. Integrate and divide the dataset: merge the initial labeled dataset and the enhanced labeled dataset. After merging, adopt a hierarchical random partitioning strategy to split the dataset into training set and test set in a 7:3 ratio, so that the data in the enhanced labeled dataset and the data before enhancement are assigned to the same set, and control the proportion of vacant disease in the training set and test set to be consistent. S34. Using the DIA-YOLO detection model as the student model, input the training set and test set into the student model to start the training of the blank detection model: use the core training parameters of the teacher model, and introduce a learning rate decay strategy. During the training process, monitor the precision, recall and mAP@0.5 of the test set in real time; when the test set metrics do not show significant improvement for 50 consecutive iterations, stop training, save the current optimal model weights, and obtain the blank detection professional model.

7. A non-destructive testing system for underground void defects, applied to the non-destructive testing method for underground void defects as described in any one of claims 1 to 6, characterized in that, The system includes: The radar data acquisition equipment includes an antenna array, an encoder, and a radar host, with both the antenna array and the encoder connected to the radar host. The computing unit is connected to the radar host and deploys the DIA-YOLO detection model; Data storage and display equipment, including a display screen and a mobile storage device, both of which are connected to a computing unit.

Images